A knowledge graph expansion and completion method based on heuristic bionic knowledge grafting strategy

Abstract

A knowledge graph expansion and completion method of a heuristic bionic knowledge grafting strategy comprises: step one, construction of a process model of heuristic bionic knowledge grafting: inspired by horticultural grafting technology, models of knowledge scion, knowledge stock and knowledge tree in the process of bionic knowledge grafting are constructed; step two, design of a guarantee mechanism for heuristic bionic knowledge grafting: based on the construction of the above-mentioned knowledge grafting process model, four specific implementation guarantee mechanisms for bionic heuristic knowledge grafting are designed, including completeness, relevance, connectivity and reusability, so as to ensure that the knowledge entity elements of engineering field technology and methods are quickly expanded to manufacturing knowledge graph and application; step three, knowledge graph expansion and completion configuration based on natural language processing; the present invention adopts natural language processing technology to improve the quality and speed of the knowledge graph expansion and completion process, realizes the automatic expansion and completion of the knowledge graph, and enhances the decision-making knowledge service capability of the knowledge graph.

Description

Technical Field

[0001] The present invention belongs to the technical field of manufacturing knowledge graph application, and in particular relates to a knowledge graph expansion and completion method of a heuristic bionic knowledge grafting strategy. Technical Background

[0002] With the innovation of new-generation information technology, production technology and production control methods, the manufacturing industry is moving from the current digitalization and automation to intelligentization driven by data and knowledge decision-making. At the same time, the rapid development of new-generation information technology represented by the Internet of Things, big data, and artificial intelligence has provided new ideas for enterprise manufacturing process decision-making that is "driven by process data and enabled by domain knowledge". In the process of transformation and upgrading of the manufacturing industry, massive multi-source heterogeneous data and knowledge centered on the entire life cycle of products are the basis for realizing the integration of manufacturing and new-generation information technology.

[0003] In the current context of intelligent manufacturing, while manufacturing enterprises are deeply analyzing the information related to the product manufacturing process, they have introduced new advanced scientific and technological methods and theories, accumulated a large amount of manufacturing-related conceptual knowledge and models, and summarized and written a large number of specific engineering case documents, which are hidden in scientific papers, patents and other documents in the form of text. At the same time, based on the construction of data / knowledge representation, mapping, reasoning and other models related to the entire life cycle of manufacturing enterprises' products, domain knowledge graphs (KGs) have been applied to the manufacturing field in the form of knowledge question-answering systems and intelligent recommendation systems, but knowledge graphs are imperfect, and the cost of completing them by traditional manual methods is huge and inefficient.

[0004] In summary, in the process of using knowledge graphs in the manufacturing field, there is a lack of a systematic method for knowledge graph entity expansion and relationship completion, which has led to fragmentation and isolation of product-related data / knowledge of manufacturing enterprises, as well as low knowledge reuse rate. Therefore, how to mine knowledge through learning of engineering case files, realize the automatic expansion and completion of the knowledge graph of manufacturing enterprises for knowledge decision-making, and enhance the knowledge service capability of the knowledge graph is an urgent problem to be solved. Summary of the invention

[0005] In order to solve the imperfection of knowledge graph in manufacturing field and improve the knowledge service performance of domain knowledge graph, the purpose of the present invention is to propose a knowledge graph expansion and completion method based on heuristic bionic knowledge grafting strategy (HGS). Firstly, on the basis of horticultural grafting and knowledge graph expansion and completion process, deduction and structural similarity analysis, the completeness, relevance, connectivity and reusability strategy of knowledge grafting is constructed, and the defects of knowledge vector space heterogeneity and knowledge noise in the traditional knowledge grafting process based on natural language processing (NLP) are solved by using K-BERT pre-training model, so as to enhance the knowledge service performance of knowledge graph in manufacturing field.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is:

[0007] A knowledge graph expansion and completion method based on a heuristic bionic knowledge grafting strategy specifically includes the following steps:

[0008] Step 1: Construction of process model of heuristic bionic knowledge grafting

[0009] Inspired by horticultural grafting technology, the models of knowledge scion (KS), knowledge stock (KR) and knowledge tree (KnowTree) in the process of bionic knowledge grafting are constructed:

[0010] The knowledge scion (KS): Assume that a knowledge source engineering case file d consists of n sentences s i The structure d = {s1, s2, ..., s n}, where a case description sentence consists of m words / terms w i Composition s = {w1, w2, ..., w m}, if word / term w i is not a punctuation mark or qualifier, then the word / term w i Defined as a knowledge scion KS, judging by the part of speech, a knowledge scion may be a noun group or a verb group;

[0011] The knowledge rootstock (KnowRoot, KR) is defined as a concept graph G = (N, E) based on the core ontology of the domain, which is understood as a knowledge tree subgraph without knowledge scion KS, where each graph node n i Represents a knowledge concept, node n i With node n j The edge between i →n j It represents the association relationship between knowledge concept pairs mapped by node pairs. In addition, similar to a grafting rootstock system, various knowledge scions are organized according to knowledge structure and class. The hierarchical structure of knowledge concepts and relationships similar to a diverging star graph determines the adjacent relationship of given knowledge concepts, while the categories of knowledge concepts and relationships as a finite set determine the shared attributes or functions according to their semantic information.

[0012] The knowledge tree (KnowTree): is defined as a knowledge graph space based on knowledge triples {h, r, t|h, t∈ε, r∈R}, which can be expanded by combining knowledge scions and knowledge rootstocks; as a manufacturing domain knowledge graph, the knowledge tree is a knowledge graph based on an ontology model, in which each edge needs to disambiguate fixed entity and relationship vocabulary;

[0013] Task h in the knowledge tree, t∈ε={KS∪KR} represents an element of the entity node set ε, where h represents the head element in the knowledge triple, and t represents the tail element in the knowledge triple; r∈R represents a set of specific types of edges between nodes, that is, the relationship between the head and tail elements of the knowledge triple;

[0014] Step 2: Design of a guarantee mechanism for heuristic bionic knowledge grafting

[0015] Based on the above knowledge grafting process model, four specific implementation guarantee mechanisms for bionic heuristic knowledge grafting are designed, including completeness, relevance, connectivity and reusability, to ensure that the knowledge entity elements of engineering technology and methods are quickly expanded to manufacturing knowledge graphs and applications:

[0016] The completeness mechanism described above: Assume that a sentence of an engineering case text contains n relevant knowledge grafts (KS1, KS2, ... KS n ), then each knowledge scion KS i Will be labeled with the corresponding label set based on the domain ontology (Label1, Label2, ...Label n ) i ;

[0017] The relatedness mechanism: Assume that in the same sentence the ith knowledge scion KS i With the Jth Knowledge Scion KS j Semantically related, then the same knowledge graft KS i and KS j Corresponding label Label i and Label j There is also a necessary correlation between them. This correlation mechanism is described as:

[0018] The connectivity mechanism described (Connectivity): Assume that the label Label i In the manufacturing domain ontology, it is functionally similar to Label j interrelated, then no matter the knowledge scion is related to KS i and KS j Whether it comes from the same or different sentences, the corresponding knowledge scion KS must be marked i and KS j Set to connect, in the defined label pair (Label i , Label j ) in the knowledge tree, the corresponding scion pair (KS i, K.S. j ) must have a path between them. Similarly, this connectivity mechanism can be described by a logical relationship: Among them, Related(Label i , Label j )∧KS i / j ∈Label i / j ;

[0019] The reusability mechanism (Reutilization): Assuming that the knowledge graph is driven by a knowledge graft KSi query related to a knowledge decision engineering problem, how to find the relevant answer KS j , in order to realize the storage and reuse of relevant knowledge solutions in the knowledge graph, and then guide the solution decision-making in the knowledge decision-making process of manufacturing enterprises. Combining the above three mechanisms of completeness, relevance and connectivity, the mechanism of knowledge reuse can be described by logical relationship as: Assume? Knowledge Scion KS i / j It must be the starting or ending node of the path of the relevant knowledge spike element, and it must have a label. i / j It is the starting or ending node of the corresponding label chain;

[0020] Step 3: Knowledge graph expansion and completion configuration based on natural language processing

[0021] Based on the above-mentioned heuristic bionic knowledge grafting process model construction and guarantee mechanism design, a five-stage knowledge graph expansion and completion configuration process based on natural language processing is proposed. This process models and configures the knowledge graph expansion and completion technology in the manufacturing field driven by natural language processing from six aspects: source knowledge data preparation and preprocessing → knowledge scion extraction → knowledge rootstock system design → knowledge grafting unit formation and knowledge scion subspace construction → knowledge tree growth based on knowledge scion pruning → knowledge reasoning based on path logic rules, so as to enhance the knowledge service performance and decision-making ability of the knowledge graph:

[0022] The source knowledge data preparation and preprocessing described above: the input of the knowledge graph expansion and completion configuration process based on natural language processing, the source data knowledge preparation and preprocessing process, the use of crawler technology to crawl the network database to collect engineering case summary ECA text pairs (d s , d r ) as source data, d s and d r are copies of each other, where d s For knowledge scion extraction, d rIt is used for knowledge stock design, and according to the writing organization structure of the engineering case summary text, the modeling expression is: ECA = {C, P, S, R}; according to the descriptive function of the sentence in the ECA text, each sentence in the ECA text is marked with CPSR tags; long independent sentences are parsed into clauses according to the target word mark, and the target words include transition phrases, conjunctions, list items, etc.; the CoreNPL tool is used to parse the entity part of speech generated by the ECA text and the dependency relationship information between entity words;

[0023] In the engineering case summary expression model ECA: C stands for Context, which indicates the background information of the engineering case, and is used for potential motivation and cause analysis of man-machine-material-method-environment-measurement, etc. From the perspective of knowledge management, it is defined as Know-Why knowledge, that is, what is the root cause of the engineering problem; P stands for Problem, which indicates a specific engineering problem, and is used to describe the goal to be achieved by the engineering research or improvement project, and the gap between the current situation and the improvement goal. From the perspective of knowledge management, it is defined as Know-What knowledge, that is, what is the specific engineering problem; S stands for Solution, which indicates the problem solution, and is the specific method, model, and technology used to solve the problem by analyzing the potential factors and root causes of the engineering problem. From the perspective of knowledge management, it is defined as Know-How knowledge, that is, what is the specific method to solve the engineering problem; R stands for Result, which indicates the problem solving result, and summarizes the improvement results of the entire case study, including improved manufacturing efficiency, improved product quality, reduced costs, shortened product delivery time, etc. From the perspective of knowledge management, it is defined as Know-With knowledge, that is, for a specific engineering problem, combined with the specific engineering background analysis, what is the improvement result obtained after adopting a specific solution.

[0024] The extraction of knowledge scions: in the process of expanding and completing the knowledge graph, following the completeness mechanism of knowledge tree grafting, all candidate knowledge scions need to be extracted from the case corpus, and each extracted candidate knowledge scion can be mapped to a knowledge stock system based on ontology concepts, that is, the ontology concept class of the knowledge graph;

[0025] First, for d sText extraction of knowledge scions, through the subject-predicate-object SAO, Subject-Action-Object structure network segmentation and part-of-speech tagging derived grammatical dependency relationship to extract each knowledge scion candidate element, for the hierarchical noun phrases and multi-word knowledge scions, the compound word compound or adjective modifier amod tagging method is used to extract noun candidate words or terms; for active sentence phrases, the clause "nsubj+dobj" is used to represent the active relationship between the subject noun phrase nsubj and the object dobj based on the "subject-predicate-object" structure, and the present participle "acl +dobj" modifies a nominal object dobj, where an adjective clause acl starts with a modified noun, and two dependency relationships extract knowledge scions; for passive clause phrases, the clause "nsubj+nmod+case" is used to describe the passive grammatical relationship chain, which separates the syntactic subject nsubj of the passive clause from the attribute-based noun modifier nmod through the case marker case, and "nmod+case+acl" is used to describe a noun modified by the noun modifier nmod with the adjective clause acl; based on the above SAO structure network knowledge scion term dependency analysis, a d s The text structure is represented as:

[0026] d s ={(S1, A1, O1), ..., (S n , A n , O n )}

[0027] Secondly, the SaoRank (ks i ) value evaluates the importance of the knowledge scion node. The SaoRank (ks i )The higher the value, the more important the knowledge graft is. Based on the network node centrality analysis of verbs and nouns, nodes with importance greater than the set threshold are selected as elements for constructing the knowledge graft unit triplet, which is expressed as:

[0028]

[0029] Where: In(ks i ) indicates knowledge graft ks i The in-degree of , represents a set of candidate sets pointing to it based on the SAO structural relationship; Ou(ks j ) indicates knowledge graft ks i The out-degree represents a set of candidates pointing to other knowledge scions based on the SAO structural relationship; d represents the damping factor, which is set to 0.85, indicating the relative importance of two or more knowledge scions associated with the same candidate in the previous text ranking.

[0030] The knowledge stock system design: According to ECA text d r The structure is organized as a set of CPSR sentence tags: r ={s1, s2, ..., s n}, combining semantic information and corresponding d s The document preprocessing sentence function analysis is annotated with a specific PCSR tag, then all the elements in the knowledge scion candidate set in the sentence can be divided into: background C, problem P, solution S and result R, that is, Know-Why / What / How / With four categories; assuming that knowledge concepts and attributes should be set as fixed terms rather than free variables, an umbrella-shaped hierarchical knowledge rootstock structure model can be designed based on the four categories of knowledge terms Know-Why / What / How / With; and ontology is selected as the clear specification for knowledge semantic conceptualization, and concepts are interconnected according to predefined relationships.

[0031] The formation of the knowledge grafting unit and the construction of the knowledge scion subspace are similar to the affinity problem between the target scion and the rootstock in the process of horticultural plant grafting, that is, the higher the affinity between the scion and the rootstock, the higher the grafting success rate; there is a problem of heterogeneous knowledge space in the knowledge grafting process. To address the problem of heterogeneous knowledge space, the K-BERT model is used to inject the knowledge triples of the original target expansion and completion knowledge graph into the preprocessed model input sentence s through knowledge injection. i ∈d r ={s1, s2, ..., s n}, and through the natural language processing technology token segmentation, position and fragment embedding and analysis, the knowledge is injected into the knowledge graft association relationship in the preprocessed sentence, the parameters in the BERT model are fine-tuned, and finally the classification task of the knowledge stock system model is set to construct the knowledge subspace based on the ontology knowledge stock CPSR. The formation of the above knowledge grafting unit and the specific implementation algorithm of the knowledge graft subspace construction;

[0032] The heterogeneous knowledge space is defined as the problem of vector heterogeneity in which the embedding vectors of the knowledge grafts in the ECA text are inconsistent with the vectors of the entities in the knowledge stock ontology constructed in a separate artificial way in the vector space dimension.

[0033] The knowledge injection is a process of injecting domain knowledge triples into the model input statement. Specifically, the process starts with the model input statement s = {w1, w2, ... i …, w n} as the benchmark, taking each knowledge graft W={w i ∈s|s∈d r ∩SaoRank(wi )≥θ} is the domain knowledge graph retrieval clue input, according to the knowledge query module Q=Query(w i , KG) in the target expansion and completion knowledge graph KG retrieval and knowledge grafting w i The corresponding knowledge triple (w i , r ij , w j ) set, and inject the set into the original benchmark sentence to form a sentence tree S_Tree = {w1, w2, ... w i {(r i0 , w i0 ),…(r ij , w ij )}…,w n}; Each triple can be regarded as a knowledge grafting unit (Union Combination), then the unit set can be expressed as: E = {(w i , r i0 , w i0 ),…,(w i , r ij , w ij )};

[0034] The knowledge subspace (KnowSpace) is defined as four types of knowledge management subset spaces of knowledge graft terms based on CPSR text labels, expressed as:

[0035] KnowSpace={KnowWhat, KnowWhy, KnowHow, KnowWith}

[0036] In the formula: KnowWhat: represents the knowledge scion term subspace that describes engineering problems; KnowWhy: represents the knowledge scion term subspace that describes engineering background; KnowHow: represents the knowledge scion term subspace that describes solutions; KnowWith: represents the knowledge scion term subspace that describes problem-solving results.

[0037] The knowledge tree growth based on knowledge scion pruning includes pruning meaningless and similar or identical knowledge noise,

[0038] In view of the meaningless knowledge noise, the feature-based Shannon entropy is first used to quantify the knowledge scion terms, and the meaningless knowledge scions that are less than the set threshold of E(ks) are deleted, which is expressed as:

[0039]

[0040] Where: P(x) represents the probability that the phrase character x belongs to the knowledge scion subspace based on the knowledge rootstock ontology model;

[0041] Secondly, the connection factor CF is designed to evaluate the relationship between each knowledge scion pair (ks i ,ks j ) and remove the second element ks in the knowledge scion pair whose value is greater than the threshold set by CF j , keep the first element ks i , the meaningless and redundant knowledge scions of the knowledge tree after simplification and expansion are expressed as:

[0042] CF(i,j)=CS(i,j)+DS(i,j)+SS(i,j)

[0043] Where: CF: represents the connectivity factor, and the larger the CF, the greater the knowledge scion pair (ks i ,ks j ) The higher the similarity between them; CS: Context Similarity, which is used to measure the knowledge graft pair (ks i ,ks j )’s text context distance; DS: distance similarity, used to measure the knowledge graft pair (ks i ,ks j ) is the path distance between the subspaces to which it belongs.

[0044]

[0045] Where: CS(i, j)∈[0,1], and the value 0 indicates that the two are not in the same text; D(i, j) represents the knowledge graft pair (ks i ,ks j ); L(text) represents the length of the text;

[0046]

[0047]

[0048] Where: DS(i, j) takes the value 0, which means that the two are not in the same sub-knowledge space; Sim C ∈[0, 1]: indicates the knowledge graft ks in the same sub-knowledge space i The concept T(i) and ks j The path distance between the concepts T(j);

[0049] SS(i, j) = con_sim(i, j)

[0050] Where: SS(i, j)∈[0,1]: represents the identification of scion pairs (ks i ,ks j) based on cosine semantic similarity con_sim(i, j), where the value 1 indicates that the semantics of the two are exactly the same;

[0051] Finally, based on the elimination and pruning of the above knowledge noise, an algorithm for knowledge tree growth based on knowledge scion pruning is designed. The algorithm process specifically includes:

[0052] Step 1: Using the knowledge triple grafting unit constructed by the knowledge injection process of the K-BERT model as a link, the knowledge scion in the sub-knowledge space is reversely mapped to the knowledge stock system of the target knowledge graph, and a weighted directed graph knowledge tree is constructed to achieve the expansion and completion of the domain knowledge graph, which can be formally expressed as:

[0053]

[0054] Where: SubSpace CPSR : represents the four knowledge scion subspaces constructed based on the CPSR knowledge scion ontology model; represents the knowledge triple grafting unit based on knowledge injection;

[0055] Step 2: Use the connectivity factor CF to quantify the triple grafting units of knowledge injection, and use the SaoRank value-based knowledge graft relevance Relatedness KS Directly mapped to the knowledge tree KnowTree node importance KnowTree The properties of the knowledge scion are ignored, and isolated nodes are ignored; at the same time, the in-and-out degree of the knowledge scion determines the direction of the node in the directed graph of the knowledge tree, which can be expressed as:

[0056] Map: Relatedness Ks →Importance KnowTree

[0057] Step 3: According to the connectivity analysis, the SaoRank value determines the priority of grafting knowledge scions into the knowledge tree. The higher the value, the higher the probability of grafting into the knowledge tree; vice versa. Select the knowledge scion candidate set KS whose SaoRank value is greater than the threshold in the knowledge grafting process o , expressed as:

[0058] KS o ={ks i , i = 1, ..., n|SaoRank(ks i )≥θ}

[0059] In the formula, ks i : represents the knowledge scion whose SaoRank value is greater than the set threshold θ in the knowledge scion candidate set; n: represents the size of the set;

[0060] Step 4: The knowledge graft elements ks whose SaoRank values ​​are greater than the set threshold θ are selected. i As an element of the triplet connecting the knowledge subspace and the extended knowledge graph, it is reversely mapped to the target extended knowledge graph and the mapping rules are designed, which can be expressed as:

[0061] RuleMa p ={ks i →kr i |ks i ∈KS o , kr i ∈KR CPSR}

[0062] In the formula, ks i ∈KS o : represents the knowledge grafting rootstock KR in the target extended knowledge graph CPSR elements in the set, the triple grafting unit is represented as (ks.r, kr)∈E, where E represents the set of knowledge grafting units.

[0063] The knowledge reasoning based on path logic rules is expressed as follows:

[0064] LR = {Map C →Map P →Map S →Map R}

[0065] Where: Map C : Indicates that given the engineering background C, the solution-type knowledge set S can be obtained by reasoning; Map P :Indicates that based on the analysis of manufacturing efficiency and quality problems and challenges P, the corresponding solutions can be obtained through Map S Mapping reasoning R : indicates that the problem solving result R is derived from the final implementation plan combined with the engineering problem and its background;

[0066] Based on the above main reasoning logic path, five sub-combination reasoning rules are designed:

[0067] R_1: Map C →Map P

[0068] Where: R_1: represents the mapping from specific engineering background knowledge to corresponding problem knowledge;

[0069] R_2: Map P →Map s

[0070] Where: R_2: represents the mapping from a specific engineering problem to a specific solution;

[0071] R_3: Map S →Map P

[0072] Where: R_3: represents the mapping from a specific solution to solving a specific type of engineering problem;

[0073] R_4:(Map C , Map P )→Map S

[0074] Where: R_4: indicates that the engineering problem can be mapped to the corresponding solution only when combined with a specific background;

[0075] R_5:(Map P , Map s )→Map R

[0076] Where: R_1: represents a given engineering problem, which can only be mapped to the corresponding solution by combining a specific solution.

[0077] The beneficial effects of the present invention are:

[0078] (1) Compared with the traditional artificial knowledge graph completion method, the bionic-inspired knowledge grafting strategy constructed by the present invention for the expansion and completion of the knowledge graph uses natural language processing technology to improve the quality and speed of the knowledge graph expansion and completion process, realizes the automatic expansion and completion of the knowledge graph, and enhances the decision-making knowledge service capability of the knowledge graph.

[0079] (2) The heuristic bionic knowledge graph expansion and completion method proposed in the present invention takes engineering case documents as data sources, and uses the SAO structure network and SaoRank network node centrality evaluation algorithm to construct a knowledge scion set; based on the analysis of the organizational structure of engineering case documents, the CPSR ontology model is used to design the knowledge stock model; the connectivity factor between the knowledge scion and the knowledge stock is set, and the K-BERT pre-trained model is used to extract manufacturing knowledge triples to realize knowledge grafting; a path-based reasoning rule is designed to verify the rationality of knowledge grafting and the reusability of knowledge. BRIEF DESCRIPTION OF THE DRAWINGS

[0080] Figure 1 It is a knowledge grafting process model based on horticultural grafting heuristic bionics.

[0081] Figure 2 It is the guarantee mechanism for grafting heuristic bionic knowledge.

[0082] Figure 3Configure the process for bionic heuristic knowledge grafting based on NLP technology.

[0083] Figure 4 Knowledge preparation and preprocessing of source data.

[0084] Figure 5 The parsing process is based on the target word token.

[0085] Figure 6 For dependency-based knowledge scion extraction.

[0086] Figure 7 It is a network node centrality analysis based on verbs and nouns.

[0087] Figure 8 Design of CPSR knowledge stock system model based on ontology model.

[0088] Fig. 9 Rootstock concept definition for ontology-based knowledge grafting.

[0089] Fig.10 Rootstock data instance relational definition for ontology-based knowledge grafting.

[0090] Fig.11 It represents the knowledge grafting process based on domain knowledge injection.

[0091] Fig.12 K-BERT model and related parameter settings based on domain knowledge injection.

[0092] Fig.13 Constructing algorithms for K-BERT-based knowledge scion subspace.

[0093] Fig.14 To grow the knowledge tree and prune the knowledge scion based on the grafting unit. DETAILED DESCRIPTION

[0094] The present invention is described in detail below with reference to the accompanying drawings and examples. The accompanying drawings described herein are a part of this application and are used to further explain the present invention, but do not constitute a limitation of the present invention.

[0095] The specific steps of the present invention include the construction of a heuristic bionic knowledge grafting process model and the design of a guarantee mechanism, as well as a knowledge graph expansion and completion configuration process based on natural language processing. The specific steps are as follows:

[0096] A knowledge graph expansion and completion method based on a heuristic bionic knowledge grafting strategy specifically includes the following steps:

[0097] Step 1: Construction of process model of heuristic bionic knowledge grafting

[0098] Inspired by horticultural grafting technology, the models of knowledge scion (KS), knowledge rootstock (KR) and knowledge tree (KnowTree) in the process of bionic knowledge grafting are constructed, see Figure 1 .

[0099] The knowledge scion (KS) is: Assume that a knowledge source engineering case file d consists of n sentences s i The structure d = {s1, s2, ..., s n}, where a case description sentence consists of m words / terms w i Composition s = {w1, w2, ..., w m}, if word / term w i is not a punctuation mark or a qualifier (e.g., a, this, that), then the word / term w i It is defined as a knowledge scion KS. Judging from the part of speech, a knowledge scion may be a noun group or a verb group.

[0100] The knowledge rootstock (KnowRoot, KR) is defined as a concept graph G = (N, E) based on the core ontology of the domain, which can be understood as a knowledge tree subgraph without the knowledge scion KS. i Represents a knowledge concept, node n i With node n j The edge between i →n j It represents the association relationship between knowledge concept pairs mapped by node pairs. In addition, similar to a grafting rootstock system, various knowledge scions are organized according to knowledge structure and class, and the hierarchical structure of knowledge concepts and relationships similar to a diverging star graph determines the adjacent relationship of a given knowledge concept, while the categories of knowledge concepts and relationships as a finite set determine the shared attributes or functions according to their semantic information.

[0101] The knowledge tree (KnowTree) is defined as a knowledge graph space based on knowledge triples {h, r, t|h, t∈ε, r∈R}, which can be expanded by combining knowledge scions and knowledge rootstocks. As a manufacturing domain knowledge graph, the knowledge tree is a knowledge graph based on an ontology model, in which each edge (i.e., entity relationship) needs to be disambiguated for a fixed entity and relationship vocabulary.

[0102] h, t∈ε={KS∪KR} in the knowledge triple represents the elements of the entity node set ε, where h represents the head element in the knowledge triple and t represents the tail element in the knowledge triple; r∈R represents a set of specific types of edges between nodes, that is, the relationship between the head and tail elements of the knowledge triple.

[0103] Step 2: Design of a guarantee mechanism for heuristic bionic knowledge grafting

[0104] Based on the above knowledge grafting process model, four specific implementation guarantee mechanisms for bionic heuristic knowledge grafting are designed, including completeness, relevance, connectivity and reusability, see Figure 2 , to ensure that the knowledge entity elements of new technologies and methods in the engineering field are quickly expanded to the manufacturing knowledge graph and application.

[0105] The completeness mechanism described above: Assume that a sentence of an engineering case text contains n relevant knowledge grafts (KS1, KS2, ... KS n ), then each knowledge scion KS i Will be labeled with the corresponding label set based on the domain ontology (Label1, Label2, ...Label n ) i The completeness mechanism ensures that each knowledge graft in the engineering case text will become a subset of the extended candidate grafts of the knowledge tree, ensuring the completeness of the heuristic bionic knowledge grafting driven knowledge graph expansion and completion proposed in the present invention.

[0106] The relatedness mechanism: Assume that in the same sentence the ith knowledge scion KS i With the Jth Knowledge Scion KS j Semantically related, then the same knowledge graft KS i and KS j Corresponding label Label i and Label j There is also a necessary correlation between them. This correlation mechanism can be described as: The relevance mechanism ensures that there are no isolated knowledge elements in the knowledge tree, and ensures the relevance of the heuristic bionic knowledge grafting driven knowledge graph expansion and completion proposed in the present invention.

[0107] The connectivity mechanism described (Connectivity): Assume that the label Label i In the manufacturing domain ontology, it is functionally similar to Label j interrelated, then no matter the knowledge scion is related to KS i and KS j Whether it comes from the same or different sentences, the corresponding knowledge scion KS must be marked i and KS j In other words, in the defined label pair (Label i , Label j ) in the knowledge tree, the corresponding scion pair (KS i , K.S.j ) must have a path between them. Similarly, this connectivity mechanism can be described by a logical relationship as: Among them, Related(Label i , Label j )∧KS i / j ∈Label i / j The connectivity mechanism ensures that there is at least one associated path between pairs of knowledge elements based on related tags, ensuring the connectivity of the heuristic bionic knowledge grafting driven knowledge graph expansion and completion proposed in the present invention.

[0108] The reusability mechanism (Reutilization): Assume that in the knowledge graph, there is a knowledge graft KS related to a knowledge decision engineering problem i Query driven, how to find relevant answers KS j , so as to realize the storage and reuse of relevant knowledge solutions in the knowledge graph, and then guide the solution decision-making in the knowledge decision-making process of manufacturing enterprises. Combining the above three mechanisms of completeness, relevance and connectivity, the mechanism of knowledge reuse can be described by logical relationship as follows: Assume? Knowledge Scion KS i / j It must be the starting or ending node of the path of the relevant knowledge spike element, and it must have a label. i / j It is the starting or ending node of the corresponding label chain. The reusability mechanism ensures that the knowledge elements extracted and stored in the knowledge tree can form knowledge solutions to guide new engineering problems through knowledge calculation and reasoning, ensuring the reusability of the heuristic bionic knowledge grafting driven knowledge graph expansion and completion proposed in this invention.

[0109] Step 1: Knowledge graph expansion and completion configuration based on natural language processing

[0110] Based on the above heuristic bionic knowledge grafting process model construction and guarantee mechanism design, a five-stage knowledge graph expansion and completion configuration process based on natural language processing is proposed, see Figure 3 The process models and configures the expansion and completion technology of the manufacturing field knowledge graph driven by natural language processing from six aspects: source knowledge data preparation and preprocessing → knowledge scion extraction → knowledge stock system design → knowledge grafting unit formation and knowledge scion subspace construction → knowledge tree growth based on knowledge scion pruning → knowledge reasoning based on path logic rules, so as to enhance the knowledge service performance and decision-making ability of the knowledge graph.

[0111] 1) Source knowledge data preparation and preprocessing: It is the input of the knowledge graph expansion and completion configuration process based on natural language processing. Source data knowledge preparation and preprocessing process, see Figure 4The present invention uses crawler technology to crawl the network database to collect engineering case summary ECA text pairs (d s , d r ) as source data (d s and d r are copies of each other, where d s For knowledge scion extraction, d r For knowledge stock design), and based on the writing organizational structure of the engineering case summary text, the modeling expression is: ECA = {C, P, S, R}; according to the statement function of the sentence in the ECA text, each sentence in the ECA text is marked with a CPSR tag; the long independent sentence is parsed into clauses according to the target word token, and the target words include transition phrases, conjunctions, list items, etc., see Figure 5 ; Use CoreNPL tool to parse ECA text to generate entity parts of speech and dependency information between entity words.

[0112] In the engineering case summary expression model ECA: C (Context) represents the background information of the engineering case, which is used for potential motivation and cause analysis of man-machine-material-method-environment-measurement, etc. From the perspective of knowledge management, it is defined as Know-Why knowledge, that is, what is the root cause of the engineering problem; P (Problem) represents the specific engineering problem, which is used to describe the goal to be achieved by the engineering research or improvement project, and the gap between the current situation and the improvement goal. From the perspective of knowledge management, it is defined as Know-What knowledge, that is, what is the specific engineering problem; S (Solution) represents the problem solution, which is the specific method, model, technology, etc. adopted by analyzing the potential factors and root cause analysis of the above engineering problems. From the perspective of knowledge management, it is defined as Know-How knowledge, that is, what is the specific method to solve the engineering problem; R (Result) represents the problem solving result, which summarizes the improvement results of the entire case study, including manufacturing efficiency improvement, product quality improvement, cost reduction, product delivery time reduction, etc. From the perspective of knowledge management, it is defined as Know-With knowledge, that is, for a specific engineering problem, combined with the specific engineering background analysis, what is the improvement result obtained after adopting a specific solution.

[0113] 2) Knowledge scion extraction: In the process of expanding and completing the knowledge graph, following the completeness mechanism of knowledge tree grafting, all candidate knowledge scions need to be extracted from the case corpus, and each extracted candidate knowledge scion can be mapped to a knowledge stock system based on ontology concepts, that is, the ontology concept class of the knowledge graph.

[0114] First, for d SThe present invention first extracts each knowledge scion candidate element through the grammatical dependency relationship derived from the subject-verb-object SAO (Subject-Action-Object) structure network segmentation and part-of-speech tagging, see Figure 6 For hierarchical noun phrases and multi-word knowledge scions, the noun candidate words or terms are extracted based on compound or adjective modifier (amod) annotation. For active sentence phrases, the clause "nsubj+dobj" is used to represent the active relationship between the subject noun phrase (nsubj) and the pointed object (dobj) based on the "subject-predicate-object" structure, and the present participle "acl+dobj" modifies a nominal object (dobj), where an adjective clause (acl) starts with a modified noun, to extract knowledge scions. For passive clause phrases, the clause "nsubj+nmod+case" is used to describe the passive grammatical relationship chain, which separates the syntactic subject (nsubj) of the passive clause from the attribute-based noun modifier (nmod) through the case marker (case), and "nmod+case+acl" is used to describe a noun modified by a noun modifier (nmod) with an adjective clause (acl). Based on the above SAO structure network knowledge graft term dependency parsing, a d s The text can be structured as:

[0115] d s ={(S1, A1, O1), ..., (S n , A n , O n )}

[0116] Secondly, the SaoRank (ks i ) value evaluates the importance of the knowledge scion node. The SaoRank (ks i )The higher the value, the more important the knowledge graft is. Based on the network node centrality analysis of verbs and nouns, see Figure 7 Based on this, nodes whose importance is greater than the set threshold are selected as elements to construct the knowledge grafting unit triples, which can be expressed as:

[0117]

[0118] Where: In(ks i ) represents the knowledge graft kS i The in-degree of , represents a set of candidate sets pointing to it based on the SAO structural relationship; Ou(ks j ) represents the knowledge graft kS iThe out-degree represents a set of candidates pointing to other knowledge scions based on the SAO structural relationship; d represents the damping factor, which is set to 0.85, indicating the relative importance of two or more knowledge scions being associated (in / out) to the same candidate in the previous text ranking.

[0119] 3) Knowledge Stock System Design: According to ECA text r The structure is organized as a set of CPSR sentence tags: r ={s1, s2, ..., s n}, combining semantic information and corresponding d s Document preprocessing sentence function analysis is annotated with a specific PCSR tag, then all the elements in the knowledge scion candidate set in the sentence can be divided into: background C, problem P, solution S and result R, that is, Know-Why / What / How / With four categories. Assuming that knowledge concepts and attributes should be set as fixed terms rather than free variables, an umbrella-shaped hierarchical knowledge rootstock structure model can be designed based on the four categories of knowledge terms Know-Why / What / How / With, see Figure 8 , and choose ontology as an explicit specification for the semantic conceptualization of knowledge, cf. Fig. 9 , and the concepts are interconnected according to predefined relationships, see Fig.10 .

[0120] 4) Formation of knowledge grafting units and construction of knowledge scion subspace: Similar to the affinity problem between target scions and rootstocks in the process of horticultural plant grafting, the higher the affinity between scions and rootstocks, the higher the success rate of grafting. Similarly, there is a problem of heterogeneous knowledge space in the process of knowledge grafting. For the problem of heterogeneous knowledge space, see Fig.11 The present invention adopts the K-BERT model to inject the knowledge triples of the original target expansion and completion knowledge graph into the preprocessed model input sentence s through knowledge injection i ∈d r ={s1, s2, ..., s n}, and inject knowledge and knowledge grafts in preprocessed sentences through natural language processing technology token segmentation, position and fragment embedding and analysis, and fine-tune the parameters in the BERT model, see Fig.12 Finally, the classification task of the knowledge rootstock system model is set to construct the knowledge subspace based on the ontology knowledge rootstock CPSR. For the specific implementation algorithm of the formation of the above knowledge grafting unit and the construction of the knowledge grafting subspace, see Fig.13 shown.

[0121] The heterogeneous knowledge space is defined as the problem of vector heterogeneity in which the embedding vectors of the knowledge grafts in the ECA text are inconsistent with the vectors of the entities in the knowledge stock ontology constructed in a separate artificial way in the vector space dimension.

[0122] The knowledge injection is a process of injecting domain knowledge triples into the model input statement. Specifically, the process is based on the model input statement s = {w1, w2, ... w i …, w n} as the benchmark, taking each knowledge graft W={w i ∈s|s∈d r ∩SaoRank(w i )≥θ} is the domain knowledge graph retrieval clue input, according to the knowledge query module Q=Query(w i , KG) in the target expansion and completion knowledge graph KG retrieval and knowledge grafting w i The corresponding knowledge triple (w i , r ij , w j ) set, and inject the set into the original benchmark sentence to form a sentence tree S_Tree = {w1, w2, ... w i {(r i0 , w i0 ),…(r ij , w ij )}…,w n Each triple can be regarded as a knowledge grafting unit (Union Combination), and the unit set can be expressed as: E = {(w i , r i0 , w i0 ),…,(w i , r ij , w ij )}.

[0123] The knowledge subspace (KnowSpace) is defined as four types of knowledge management subset spaces of knowledge graft terms based on CPSR text labels, which can be expressed as:

[0124] KnowSpace={KnowWhat, KnowWhy, KnowHow, KnowWith}

[0125] In the formula: KnowWhat: represents the knowledge scion term subspace that describes engineering problems; KnowWhy: represents the knowledge scion term subspace that describes engineering background; KnowHow: represents the knowledge scion term subspace that describes solutions; KnowWith: represents the knowledge scion term subspace that describes problem-solving results.

[0126] 5) Knowledge tree growth based on knowledge scion pruning: mainly includes pruning meaningless and similar or identical knowledge noise, otherwise it will cause concept drift of data and knowledge elements, that is, the concept of knowledge terminology is ambiguous and there is data redundancy in the knowledge graph after expansion and completion. Although most short noun knowledge scion entities can be identified in the process of knowledge scion extraction and classification based on K-BERT, there is a lack of effective methods to eliminate knowledge noise.

[0127] In view of meaningless knowledge noise, the present invention firstly quantifies the knowledge graft terms by using the feature-based Shannon entropy, and deletes the meaningless knowledge grafts that are less than the set threshold of E(ks), which can be expressed as:

[0128]

[0129] Where: P(x) represents the probability that the phrase character x belongs to the knowledge scion subspace based on the knowledge rootstock ontology model.

[0130] Secondly, the connection factor CF is designed to evaluate the relationship between each knowledge scion pair (ks i ,ks j ) and remove the second element ks in the knowledge scion pair whose value is greater than the threshold set by CF j , keep the first element ks i , the meaningless and redundant knowledge scions of the knowledge tree after streamlining and expansion can be expressed as:

[0131] CF(i,j)=CS(i,j)+DS(i,j)+SS(i,j)

[0132] Where: CF: represents the connectivity factor, and the larger the CF, the greater the knowledge scion pair (ks i ,ks j ) The higher the similarity between them; CS: stands for context similarity, which is used to measure the knowledge scion pair (ks i ,ks j ) of the text context distance; DS: represents the path similarity (Distance Similarity), which is used to measure the knowledge scion pair (ks i ,ks j ) is the path distance between the subspaces to which it belongs.

[0133]

[0134] Where: CS(i, j)∈[0,1], and the value 0 indicates that the two are not in the same text; D(i, j) represents the knowledge graft pair (ks i ,ks j );L(text) represents the length of the text.

[0135]

[0136]

[0137] Where: DS(i, j) takes the value 0, which means that the two are not in the same sub-knowledge space; Sim C ∈[0, 1]: indicates the knowledge graft ks in the same sub-knowledge space i The concept T(i) and ks j The path distance between the concepts T(j);

[0138] SS(i, j) = con_sim(i, j)

[0139] Where: SS(i, j)∈[0,1]: represents the identification of scion pairs (ks i ,ks j ) based on cosine semantic similarity con_sim(i, j), where the value 1 indicates that the semantics of the two are exactly the same.

[0140] Finally, based on the elimination and pruning of the above-mentioned knowledge noise, the present invention designs an algorithm for knowledge tree growth based on knowledge scion pruning, see Fig.14 The algorithm process specifically includes:

[0141] Step 1: Using the knowledge triple grafting unit constructed by the knowledge injection process of the K-BERT model as a link, the knowledge scion in the sub-knowledge space is reversely mapped to the knowledge stock system of the target knowledge graph, and a weighted directed graph knowledge tree is constructed to achieve the expansion and completion of the domain knowledge graph, which can be formally expressed as:

[0142]

[0143] Where: SubSpace CPSR : represents the four knowledge scion subspaces constructed based on the CPSR knowledge scion ontology model; Represents the knowledge triple grafting unit based on knowledge injection.

[0144] Step 2: Use the connectivity factor CF to quantify the triple grafting units of knowledge injection, and use the SaoRank value-based knowledge graft relevance RelatednessKS Directly mapped to the knowledge tree KnowTree node importance KnowTree The properties of the knowledge scion are ignored, and isolated nodes are ignored. At the same time, the in-and-out degrees of the knowledge scion determine the direction of the node in the directed graph of the knowledge tree. It can be expressed as:

[0145] Map: Relatedness KS →Importance KnowTree

[0146] Step 3: According to the connectivity analysis, the SaoRank value determines the priority of grafting the knowledge scion to the knowledge tree. The higher the value, the higher the probability of grafting to the knowledge tree; otherwise, the lower the probability. Specifically, the present invention selects the knowledge scion candidate set KS whose SaoRank value is greater than the threshold value in the knowledge grafting process. o , which can be expressed as:

[0147] KS o ={ks i , i = 1, ..., n|SaoRank(ks i )≥θ}

[0148] In the formula, ks i : represents the knowledge scion whose SaoRank value is greater than the set threshold θ in the knowledge scion candidate set; n: represents the size of the set.

[0149] Step 4: The knowledge graft elements ks whose SaoRank values ​​are greater than the set threshold θ are selected. i As an element of the triplet connecting the knowledge subspace and the extended knowledge graph, it is reversely mapped to the target extended knowledge graph and the mapping rules are designed, which can be expressed as:

[0150] Rule Map ={ks i →kr i |ks i ∈KS o , kr i ∈KR CPSR}

[0151] In the formula, ks i ∈KS o : represents the knowledge grafting rootstock KR in the target extended knowledge graph CPSR elements in the set, the triple grafting unit can be expressed as (ks.r, kr)∈E, where E represents the set of knowledge grafting units.

[0152] 6) Knowledge reasoning based on path logic rules: In order to realize the reasoning of sub-knowledge space in the expanded and completed knowledge tree and the reusability of knowledge grafts, the present invention proposes a reasoning rule based on path logic, which can be expressed as:

[0153] LR = {Map C →Map p →Map s →Map R}

[0154] Where: Map C : Indicates that given the engineering background C, the solution-type knowledge set S can be obtained by reasoning; Map P :Indicates that based on the analysis of manufacturing efficiency and quality problems and challenges P, the corresponding solutions can be obtained through Map S Mapping reasoning R : It indicates that the problem-solving result R is derived from the final implementation plan combined with the engineering problem and its background.

[0155] Based on the above main reasoning logic path, the present invention designs five sub-combination reasoning rules:

[0156] R_1: Map C →Map P

[0157] Where: R_1: represents the mapping from specific engineering background knowledge to corresponding problem knowledge;

[0158] R_2: Map P →Map S

[0159] Where: R_2: represents the mapping from a specific engineering problem to a specific solution;

[0160] R_3: Map S →Map P

[0161] Where: R_3: represents the mapping from a specific solution to solving a specific type of engineering problem;

[0162] R_4:(Map C , Map P )→Map S

[0163] Where: R_4: indicates that the engineering problem can be mapped to the corresponding solution only when combined with a specific background;

[0164] R_5:(Map P , Map S )→Map R Where: R_1: represents a given engineering problem, which can only be mapped to the corresponding solution by combining a specific solution.

Claims

1. A knowledge graph expansion and completion method based on a heuristic bionic knowledge grafting strategy, characterized in that: The specific steps include: Step 1: Construction of process model of heuristic bionic knowledge grafting Inspired by horticultural grafting technology, the models of knowledge scion, knowledge rootstock and knowledge tree in the process of bionic knowledge grafting are constructed; Step 2: Design of a guarantee mechanism for heuristic bionic knowledge grafting Based on the above knowledge grafting process model, four specific implementation guarantee mechanisms for bionic heuristic knowledge grafting are designed, including completeness, relevance, connectivity and reusability, to ensure that the knowledge entity elements of engineering technology and methods are quickly expanded to manufacturing knowledge graphs and applications; Step 3: Knowledge graph expansion and completion configuration based on natural language processing Based on the above-mentioned heuristic bionic knowledge grafting process model construction and guarantee mechanism design, a five-stage knowledge graph expansion and completion configuration process based on natural language processing is proposed. This process models and configures the knowledge graph expansion and completion technology in the manufacturing field driven by natural language processing from six aspects: source knowledge data preparation and preprocessing → knowledge scion extraction → knowledge rootstock system design → knowledge grafting unit formation and knowledge scion subspace construction → knowledge tree growth based on knowledge scion pruning → knowledge reasoning based on path logic rules, so as to enhance the knowledge service performance and decision-making ability of the knowledge graph; In the step 2: The completeness mechanism described: Assume that a sentence of an engineering case text contains n relevant knowledge grafts (KS1, KS2, ... KS n ), then each knowledge scion KS i Will be labeled with the corresponding label set based on the domain ontology (Label1, Label2, ...Label n ) i ; The relatedness mechanism: Assume that in the same sentence the ith knowledge scion KS i With the Jth Knowledge Scion KS j Semantically related, then the same knowledge graft KS i and KS j Corresponding label Label i and Label j There is also a necessary correlation between them. This correlation mechanism is described as: The connectivity mechanism described (Connectivity): Assume that the label Label i In the manufacturing domain ontology, it is functionally similar to Label j interrelated, then no matter the knowledge scion is related to KS i and KS j Whether it comes from the same or different sentences, the corresponding knowledge scion KS must be marked i and KS j Set to connect, in the defined label pair (Label i ,Label j ) in the knowledge tree, the corresponding scion pair (KS i ,KS j ) must have a path between them. Similarly, this connectivity mechanism can be described by a logical relationship: Among them, Related(Label i ,Label j )∧KS i / j ∈Label i / j ; The reusability mechanism (Reutilization): Assume that in the knowledge graph, there is a knowledge graft KS related to a knowledge decision engineering problem i Query driven, how to find relevant answers KS j In order to realize the storage and reuse of relevant knowledge solutions in the knowledge graph, and then guide the solution decision-making in the knowledge decision-making process of manufacturing enterprises, combined with the above three mechanisms of completeness, relevance and connectivity, the mechanism of knowledge reuse can be described by logical relationship as follows: Assume Knowledge Scion KS i / j It must be the starting or ending node of the path of the relevant knowledge spike element, and it must have a label. i / j It is the start or end node of the corresponding label chain.

2. The knowledge graph expansion and completion method of the heuristic bionic knowledge grafting strategy according to claim 1 is characterized in that: In the step 1: The knowledge graft: Assume that a knowledge source engineering case file d consists of n sentences s i The structure d={s1,s2,…,s n }, where a case description sentence consists of m words / terms w i Composition s={w1,w2,…,w m }, if word / term w i If it is not a punctuation mark or a qualifier, then the word / term wi is defined as a knowledge scion KS. According to the part of speech, a knowledge scion may be a noun group or a verb group; The knowledge rootstock is defined as a concept graph G = (N, E) based on the core ontology of the domain, which is understood as a knowledge tree subgraph without knowledge scion KS, where each graph node n i Represents a knowledge concept, node n i With node n j The edge between i →n j It represents the association relationship between knowledge concept pairs mapped by node pairs. In addition, similar to a grafting rootstock system, various knowledge scions are organized according to knowledge structure and class. The hierarchical structure of knowledge concepts and relationships similar to a diverging star graph determines the adjacent relationship of given knowledge concepts, while the categories of knowledge concepts and relationships as a finite set determine the shared attributes or functions according to their semantic information. The knowledge tree is defined as a knowledge graph space based on knowledge triples {h, r, t|h, t∈ε, r∈R}, which can be extended by combining knowledge scions and knowledge rootstocks; as a manufacturing domain knowledge graph, the knowledge tree is a knowledge graph based on an ontology model, in which each edge needs to disambiguate fixed entity and relationship vocabulary; The task h,t∈ε={KS∪KR} in the knowledge tree represents the elements of the entity node set ε, where h represents the head element in the knowledge triple, and t represents the tail element in the knowledge triple; r∈R represents a set of specific types of edges between nodes, that is, the relationship between the head and tail elements of the knowledge triple.

3. The knowledge graph expansion and completion method of the heuristic bionic knowledge grafting strategy according to claim 1 is characterized in that: In the step three: Source knowledge data preparation and preprocessing: Based on the input of the knowledge graph expansion and completion configuration process based on natural language processing, source data knowledge preparation and preprocessing process, using crawler technology to crawl the network database to collect engineering case summary ECA text pairs (d s ,d r ) as source data, d s and d r are copies of each other, where d s For knowledge scion extraction, d r It is used for knowledge stock design, and according to the writing organization structure of the engineering case summary text, the modeling expression is: ECA = {C, P, S, R}; according to the descriptive function of the sentence in the ECA text, each sentence in the ECA text is marked with CPSR tags; long independent sentences are parsed into clauses according to the target word mark, and the target words include transition phrases, conjunctions, and list items; the CoreNPL tool is used to parse the entity part of speech generated by the ECA text and the dependency relationship information between entity words; In the engineering case summary expression model ECA: C stands for Context, which indicates the background information of the engineering case, and is used for the potential motivation and cause analysis of man-machine-material-method-environment-measurement. From the perspective of knowledge management, it is defined as Know-Why knowledge, that is, what is the root cause of the engineering problem; P stands for Problem, which indicates a specific engineering problem, and is used to describe the goal to be achieved by the engineering research or improvement project, and the gap between the current situation and the improvement goal. From the perspective of knowledge management, it is defined as Know-What knowledge, that is, what is the specific engineering problem; S stands for Solution, which indicates the problem solution. By analyzing the potential factors and root cause analysis of the engineering problem, the specific problem-solving methods, models, and technologies adopted are defined as Know-How knowledge from the perspective of knowledge management, that is, what is the specific method to solve the engineering problem; R stands for Result, which indicates the problem-solving result, and summarizes the improvement results of the entire case study, including improved manufacturing efficiency, improved product quality, reduced costs, and shortened product delivery time; from the perspective of knowledge management, it is defined as Know-With knowledge, that is, for a specific engineering problem, combined with the specific engineering background analysis, what is the improvement result obtained after adopting a specific solution.

4. The knowledge graph expansion and completion method of the heuristic bionic knowledge grafting strategy according to claim 1 is characterized in that: In the step three: Knowledge scion extraction: In the process of expanding and completing the knowledge graph, following the completeness mechanism of knowledge tree grafting, all candidate knowledge scions need to be extracted from the case corpus, and each extracted candidate knowledge scion can be mapped to a knowledge rootstock system based on ontology concepts, that is, the ontology concept class of the knowledge graph; First, for d s Extract knowledge scions from texts. Each knowledge scion candidate element is extracted through grammatical dependencies derived from network word segmentation and part-of-speech tagging based on subject-verb-object SAO, Subject-Action-Object structure. For hierarchical noun phrases and multi-word knowledge scions, noun candidates or terms are extracted by tagging based on compound words or adjective modifiers amod. For active sentence phrases, the clause "nsubj+dobj" is used to express the active relationship between the subject noun phrase nsubj and the object dobj based on the "subject-predicate-object" structure, and the present participle "acl+dobj" modifies a nominal object dobj, where an adjective clause acl starts with a modified noun. The two dependency relationships extract knowledge scions; for passive clause phrases, the clause "nsubj+nmod+case" is used to describe the passive grammatical relationship chain, which separates the syntactic subject nsubj of the passive clause from the attribute-based noun modifier nmod through the case marker case, and "nmod+case+acl" is used to describe a noun modified by the noun modifier nmod with the adjective clause acl; based on the above SAO structure network knowledge scion term dependency analysis, a d s The text structure is represented as: d s ={(S1,A1,O1),…,(S n ,A n ,O n )} Secondly, the SaoRank (ks i ) value evaluates the importance of the knowledge scion node. The SaoRank (ks i )The higher the value, the more important the knowledge graft is. Based on the network node centrality analysis of verbs and nouns, nodes with importance greater than the set threshold are selected as elements for constructing the knowledge graft unit triplet, which is expressed as: Where: In(ks i ) indicates knowledge graft ks i The in-degree of , represents a set of candidate sets pointing to it based on the SAO structural relationship; Ou(ks j ) indicates knowledge graft ks i The out-degree represents a set of candidates pointing to other knowledge scions based on the SAO structural relationship; d represents the damping factor, which is set to 0.85, indicating the relative importance of two or more knowledge scions associated with the same candidate in the previous text ranking.

5. The knowledge graph expansion and completion method of the heuristic bionic knowledge grafting strategy according to claim 1 is characterized in that: In the step three: Knowledge Stock System Design: According to ECA Text d r The structure is organized as a set of CPSR sentence tags: r ={s1,s2,…,s n }, combining semantic information and corresponding d s The document preprocessing sentence function analysis is annotated with a specific PCSR tag, then all the elements in the knowledge scion candidate set in the sentence can be divided into: background C, problem P, solution S and result R, that is, Know-Why / What / How / With four categories; assuming that knowledge concepts and attributes should be set as fixed terms rather than free variables, an umbrella-shaped hierarchical knowledge rootstock structure model can be designed based on the four categories of knowledge terms Know-Why / What / How / With; and ontology is selected as the clear specification for knowledge semantic conceptualization, and concepts are interconnected according to predefined relationships.

6. The knowledge graph expansion and completion method of the heuristic bionic knowledge grafting strategy according to claim 1 is characterized in that: In the step three: The formation of knowledge grafting units and the construction of knowledge scion subspace: similar to the affinity problem between target scions and rootstocks in the process of horticultural plant grafting, that is, the higher the affinity between scions and rootstocks, the higher the grafting success rate; In the process of knowledge grafting, there is a problem of heterogeneous knowledge space. To address this problem, the K-BERT model is used to inject knowledge triples of the original target expansion and completion knowledge graph into the preprocessed model input sentence s. i ∈d r ={s1,s2,…,s n }, and through the natural language processing technology token segmentation, position and fragment embedding and analysis, the knowledge is injected into the knowledge scion association relationship in the preprocessed sentence, the parameters in the BERT model are fine-tuned, and finally the classification task of the knowledge rootstock system model is set to construct the knowledge subspace based on the ontology knowledge rootstock CPSR; the formation of the above-mentioned knowledge grafting unit and the specific implementation algorithm of the knowledge scion subspace construction; The heterogeneous knowledge space is defined as the inconsistency between the embedding vectors of the knowledge grafts in the ECA text and the vectors of the entities in the knowledge stock ontology constructed in a separate artificial way in the vector space dimension, i.e., the vector heterogeneity problem; The knowledge injection is a process of injecting domain knowledge triples into model input statements. Specifically, the process takes the model input sentence s = {w1, w2, ... w i …,w n } as the benchmark, taking each knowledge graft W={w i ∈s|s∈d r ∩SaoRank(w i )≥θ} is the domain knowledge graph retrieval clue input, according to the knowledge query module Q=Query(w i ,KG) in the target expansion and completion knowledge graph KG retrieval and knowledge grafting w i The corresponding knowledge triple (w i ,r ij ,w j ) set, and inject the set into the original benchmark sentence to form a sentence tree S_Tree = {w1, w2, ... w i {(r i0 ,w i0 ),…(r ij ,w ij )}…,w n }; Each triple can be regarded as a knowledge grafting unit (Union Combination), then the unit set can be expressed as: E = {(w i ,r i0 ,w i0 ),…,(w i ,r ij ,w ij )}; The knowledge subspace (KnowSpace) is defined as four types of knowledge management subset spaces of knowledge graft terms based on CPSR text labels, expressed as: KnowSpace={KnowWhat,KnowWhy,KnowHow,KnowWith} In the formula: KnowWhat: represents the knowledge scion term subspace that describes engineering problems; KnowWhy: represents the knowledge scion term subspace that describes engineering background; KnowHow: represents the knowledge scion term subspace that describes solutions; KnowWith: represents the knowledge scion term subspace that describes problem-solving results.

7. The knowledge graph expansion and completion method of the heuristic bionic knowledge grafting strategy according to claim 1 is characterized in that: In the step three: Knowledge tree growth based on knowledge scion pruning: including pruning meaningless and similar or identical knowledge noise, In view of the meaningless knowledge noise, the feature-based Shannon entropy is first used to quantify the knowledge scion terms, and the meaningless knowledge scions that are less than the set threshold of E(ks) are deleted, which is expressed as: Where: P(x) represents the probability that the phrase character x belongs to the knowledge scion subspace based on the knowledge rootstock ontology model; Secondly, the connection factor CF is designed to evaluate the relationship between each knowledge scion pair (ks i ,ks j ) and remove the second element ks in the knowledge scion pair whose value is greater than the threshold set by CF j , keep the first element ks i ,The meaningless and redundant knowledge scions of the knowledge tree after simplification and expansion are expressed as: CF(i,j)=CS(i,j)+DS(i,j)+SS(i,j) Where: CF: represents the connectivity factor, and the larger the CF, the greater the knowledge scion pair (ks i ,ks j ) the higher the similarity between them; CS: Context Similarity, used to measure the knowledge scion pair (ks i ,ks j ) of the text context distance; DS: represents the path similarity Distance Similarity, which is used to measure the knowledge scion pair (ks i ,ks j ) path distance between the subspaces to which it belongs; Where: CS(i,j)∈[0,1], and the value 0 indicates that the two are not in the same text; D(i,j) represents the knowledge graft pair (ks i ,ks j ); L(text) represents the length of the text; Where: DS(i,j) takes the value 0, which means that the two are not in the same sub-knowledge space; Sim C ∈[0,1]: indicates the knowledge graft ks in the same sub-knowledge space i The concept T(i) and ks j The path distance between the concepts T(j); SS(i,j)=con_sim(i,j) Where: SS(i,j)∈[0,1] ： Indicates identification of scion pair (ks i ,ks j ) based on cosine semantic similarity con_sim(i,j), where the value 1 indicates that the semantics of the two are exactly the same; Finally, based on the elimination and pruning of the above knowledge noise, an algorithm for knowledge tree growth based on knowledge scion pruning is designed. The algorithm process specifically includes: Step 1: Using the knowledge triple grafting unit constructed by the knowledge injection process of the K-BERT model as a link, the knowledge scion in the sub-knowledge space is reversely mapped to the knowledge stock system of the target knowledge graph, and a weighted directed graph knowledge tree is constructed to achieve the expansion and completion of the domain knowledge graph, which can be formally expressed as: Where: SubSpace CPSR : represents the four knowledge scion subspaces constructed based on the CPSR knowledge scion ontology model; represents the knowledge triple grafting unit based on knowledge injection; Step 2: Use the connectivity factor CF to quantify the triple grafting units of knowledge injection, and use the SaoRank value-based knowledge graft relevance Relatedness KS Directly mapped to the knowledge tree KnowTree node importance KnowTree The properties of the knowledge scion are ignored, and isolated nodes are ignored; at the same time, the in-and-out degree of the knowledge scion determines the direction of the node in the directed graph of the knowledge tree, which can be expressed as: Map:Relatedness KS →Importance KnowTree Step 3: According to the connectivity analysis, the SaoRank value determines the priority of grafting knowledge scions into the knowledge tree. The higher the value, the higher the probability of grafting into the knowledge tree; vice versa. Select the knowledge scion candidate set KS whose SaoRank value is greater than the threshold in the knowledge grafting process o , expressed as: KS o ={ks i ,i=1,…,n|SaoRank(ks i )≥θ} In the formula, ks i : represents the knowledge scion whose SaoRank value is greater than the set threshold θ in the knowledge scion candidate set; n: represents the size of the set; Step 4: The knowledge graft elements ks whose SaoRank values ​​are greater than the set threshold θ are selected. i As an element of the triplet connecting the knowledge subspace and the extended knowledge graph, it is reversely mapped to the target extended knowledge graph and the mapping rules are designed, which can be expressed as: Rule Map ={ks i →kr i |ks i ∈KS o ,kr i ∈KR CPSR } In the formula, ks i ∈KS o : represents the knowledge grafting rootstock KR in the target extended knowledge graph CPSR elements in the set, the triple grafting unit is represented as (ks.r,kr)∈E, where E represents the set of knowledge grafting units.

8. The knowledge graph expansion and completion method of the heuristic bionic knowledge grafting strategy according to claim 1 is characterized in that: In the step three: Knowledge reasoning based on path logic rules: expressed as: LR={Map C →Map P →Map S →Map R } Where: Map C : Indicates that given the engineering background C, the solution-type knowledge set S can be obtained by reasoning; Map P :Indicates that based on the analysis of manufacturing efficiency and quality problems and challenges P, the corresponding solutions can be obtained through Map S Mapping reasoning; Map R : indicates that the problem solving result R is derived from the final implementation plan combined with the engineering problem and its background; Based on the above path logic rules, five sub-combination reasoning rules are designed: R_1:Map C →Map P Where: R_1: represents the mapping from specific engineering background knowledge to corresponding problem knowledge; R_2:Map P →Map S Where: R_2: represents the mapping from a specific engineering problem to a specific solution; R_3:Map S →Map P Where: R_3: represents the mapping from a specific solution to solving a specific type of engineering problem; R_4:(Map C ,Map P )→Map S Where: R_4: indicates that the engineering problem can be mapped to the corresponding solution only when combined with a specific background; R_5:(Map P ,Map S )→Map R Where: R_1: represents a given engineering problem, which can only be mapped to the corresponding solution by combining a specific solution.

Images